Hardware Up 15%, Money Up to 7%, and Flat-Rate Plans Break First

What Got Expensive Is the Memory Inside the Box

Nvidia has notified Microsoft, Google, and Oracle that it will raise AI server and chip prices by more than 15%. Bloomberg reported the move on August 22. It covers systems built on Vera Rubin and Grace Blackwell, and it applies to units shipping from early 2027. The exact increase varies by product, depending on chip generation and memory configuration.

What makes this notable is that the stated reason is not demand. Nvidia points at HBM and DRAM. TrendForce counted server DRAM up 53% to 58% quarter over quarter in Q2 of this year, with another 13% to 18% forecast for Q3. Three companies, Samsung, SK Hynix, and Micron, hold the DRAM market between them, so supply does not chase a demand spike quickly.

The pattern has been running for months. Apple raised prices in June, Qualcomm in July, and now Nvidia. Some server builders were separately told AI chip prices could climb around 17%. At building scale the arithmetic compresses into one sentence: a gigawatt-class data center now costs at least $5 billion more to build.

Orders Came in Six Times Over, and the Coupon Was Still 7.228%

The other half of the cost is where the money to buy that hardware comes from and what it costs. QTS Realty, the Blackstone-backed data center operator, recently placed $3.9 billion of five-year paper at 7.228%. Moody’s rates it Baa3 and Fitch BBB-, the bottom rung of investment grade, one notch above junk.

That yield is not a demand problem. Orders reached $23 billion, roughly six times the issue, against an average of about four times for investment-grade corporates this year, and the strength let the issuer shave 0.4 percentage points off initial guidance. It still landed above 7%. Morgan Stanley’s read is that when spreads widen, the weight falls disproportionately on lower-rated borrowers. Vivek Paul of the BlackRock Investment Institute frames it as a capital shortage: AI investment keeps expanding, and the gap shows up in bond yields.

The clearest measure is the same issuer four months apart. QTS sold $4.6 billion of ten-year paper at 5.7% in April, and that bond now trades around 7.16%. The $12.55 billion bond BlackRock arranged for Meta data centers, maturing in 2048, carries 7.53%. Those three numbers bracket what money for data centers currently costs.

Count the Floor Price of a GPU Hour, Then Move It to Your Price Sheet

Put both increases on one line. Call the hardware 100 before the hike and 115 after. Add financing on top, treating the principal as repaid at maturity and simply summing the interest:

• Before: principal 100 plus interest (100 × 5.7% × 5 years) 28.5 = 128.5
• After: principal 115 plus interest (115 × 7.2% × 5 years) 41.4 = 156.4

Five-year total goes from 128.5 to 156.4, up about 22%. The hardware line moved 15%; financing carried it to 22%. This ignores depreciation schedules, power, and lease structure, so it will not match a real contract. It does get the direction and the order of magnitude right.

Divide 156.4 by the hours in five years and you get the floor price of a GPU hour. That division assumes the machine runs at full utilization for the entire period. At half utilization the hourly cost doubles. The denominator is why suppliers want long commitments and money up front.

That is the infrastructure side. The part that actually reaches a founder is where that 22% lands on their own price sheet, and the order is fairly consistent.

Flat-rate unlimited plans leak first. Revenue is fixed per month while cost tracks usage. A 22% increase leaves the average customer profitable and widens the loss on heavy ones. Margins are calculated on averages; cash leaves on totals.

Free trial tiers come second. Cost rises and revenue stays at zero, so the increase converts directly into marketing spend. If the cap on that tier is written in tokens, the line item grows on its own as unit prices move.

Per-seat pricing is third. Seats grow when headcount grows; token consumption grows when the same people use the tool more. Where those two curves separate, seat prices stop tracking cost.

Fourth are annual prepaid contracts already sold. The rate is locked at the old cost for twelve months, so there is nothing to adjust until renewal.

Three tables are worth pulling now. Take the actual consumption of your top 5% of users, multiply cost by 1.22, and check whether that plan still carries margin. Lay annual contract renewal dates against early 2027 and see how much contracted revenue passes through that window untouched, because that figure is the real exposure. Redefine the free tier cap in currency rather than tokens, so the ceiling drops by itself when unit prices rise.

What remains unsettled is how far down the increase travels. Nvidia notified Microsoft, Google, and Oracle, and none of the three has said when or how their own rate cards change. If list prices hold while only new commitment terms tighten, teams sitting on existing contracts will feel nothing until renewal. That is why the date worth watching is not a pricing announcement but the renewal date on your own agreement.